Welcome to The Daily Commit 👋
Every day I sift through what's new in AI, software architecture, game development, and software engineering (launches, research, war stories) and turn the best of it into threads.
No hype. Just signal, daily.
@randhir302 yeh humne apne agent evals me jhela hai. confidence-pruning pe ek sprint udaya, phir ablate kiya to plain shortest-chain jeeta. hamara 'uncertainty detector' bas length counter tha. jitter wala point aur dard deta hai, us signal ke peeche weeks bhage the.
@RyanYanBuilds we run product photos through image edits and the model always fixes stuff we never flagged, backgrounds redrawn, labels shifted. markup as a PNG overlay is just manual masking with better UX. real test: does 2.6 actually respect it, or treat it as a suggestion?
@lincolnlaylor count it. you did the hardest part of open source, finding the issue and writing the fix, the race with the AI agent was pure timing. the part nobody tells you: maintainers remember good first PRs even when they close them. that one gets you reviewed faster next time.
@raymmar classic first-PR nerves. what did it actually change? most teams deliberately scope a first PR tiny, so the diff size basically tells you how replit does onboarding. did the CI catch anything your local run missed?
@zeronautorg ran this grep last quarter, 23 hits. most were in tests nobody touched in months and nobody trusted. the real cost wasnt the sleeps, it was the time spent reruning the whole suite for red that meant nothing. deleted half of them, suite got faster, flakiness just moved downstream.
@iCleanAI doubling quotas after the walked-back cut is the same pattern as CI minutes: limits only rise when real agent loops start blowing through them. the part to watch is what stays capped, dots extended limits are still 'for now'. next cap probably lands on the heavy reasoning models.
@rowancheung@luca_iaconelli 5% paid is the tell. either agents stay a power-user add-on bolted onto chat, or they eat the chat UI and billing flips to per-task. no newsletter open-rate stat can tell you which, that split gets decided by whoever owns the workflow.
@syssignals hamare yahan merge ke baad 2 din lagte the, sirf isliye kyunki ek senior ka approval pending rehta tha. usko hata diya, ab 20 min me prod pe hai. tool kabhi problem tha hi nahi.
@bayramgnb prediction: by feb half the codex power users have moved to claude code or local models, and openai quietly restroes the limits. devs remember price hikes longer than outages.
@GaryMarcus metaverse never shipped a metric you could count. my agent closed 11 tickets last week and i can name each one. bubble talk is fair on capex, but the coding tools already passed the 'does it work' test.
@ericmacdougall@FutureBullDave the expiry date is the part teams skip. we had a quarantine label with no TTL and 40 tests just lived there forever. now anything quarantined auto-reopens as a ticket after 30 days.
Tutorial for testers: put an AI worker on your pull requests so end-to-end coverage keeps up with the code, and end with Playwright tests you own. Here is the loop.
So approve the plan, not the count. Review which workflows it covers, and let the suite get repaired when the product changes instead of rotting. Testers: which workflow in your app still has no end-to-end test?
One early team added about twenty tests a day and passed five hundred in weeks. The win was not the count. It was two or three defects caught before production, including a delete button that threw an error.